What This Category’s Data Looks Like
This category’s data typically comes from advertising platform backends, attribution systems, CRM lead databases, and landing page statistics tools. Update cycles are mostly daily or real-time, to meet the need for immediate adjustments to delivery performance. Daily processed documents include advertising delivery daily and weekly reports, keyword price adjustment records, material A/B test tables, and optimization review reports. Fields cover business metrics such as impressions, clicks, conversions, customer acquisition cost, and budget consumption. There are no unified fixed units, and matching must be done according to specific data types.
Constraints Imposed on Reranking and Hybrid Retrieval
The need to integrate multi-source data requires hybrid retrieval to cover different data sources such as advertising platforms, attribution systems, and CRMs, to avoid information bias from a single data source. The real-time or daily update cycle requires retrieval configurations to support incremental pulling and fast response, to avoid delays affecting optimization decisions. The diversity of business fields means reranking must take business metrics such as conversions and costs as the core; relying solely on semantic similarity cannot meet business needs. The time-consuming anomaly locating and budget allocation processes require retrieval results to accurately filter irrelevant content, reducing the workload of manual screening. The consequences of errors directly affect customer acquisition cost and budget efficiency, so the accuracy of reranking and hybrid retrieval must adapt to the rigor required for business decision-making.
How to Configure Settings
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
usingReRank | Enabled | Delivery effect optimization requires sorting by business metrics such as conversions and customer acquisition cost; basic semantic similarity cannot match business needs |
similarity | 0.65–0.75 | Filter low-relevance historical advertising data and lead records, retain valid business-related content |
limit | Top 10–15 entries | Adapt to the length of single documents such as daily reports and price adjustment records, avoid context overflow affecting model inference |
searchMode | Hybrid Retrieval | Integrate advertising platform data, attribution system records, and CRM leads to cover the complete delivery business chain |
| Reranking Model Selection | bge-rerank-base | Adapt to the semantic and business association ranking needs of advertising business, aligns with common community deployment solutions |
| Retrieval Data Source Scope | Advertising data, attribution records, and CRM leads from the past 7 days | The conventional decision cycle for delivery effect optimization is 7 days, covering the valid period for current delivery adjustments |
| FastGPT Version | FastGPT v4.17.0 | Ensure compatibility between configuration items and interface functions |
Three Common Configuration Mistakes
- Symptom: After deploying the
bge-rerank-baseimage, testing returns a 401 unauthorized error, while configurations on other platforms work normally. Cause: The correct image access key was not filled in the model configuration page of FastGPT v4.17.0, or the network permissions of the corresponding container were not enabled. - Symptom: After enabling the reranking function, the number of retrieval results is less than the preset
limitvalue. Cause: The reranking model filters results with similarity below the threshold, causing the final number of valid entries returned to be lower than the initial recall upper limit. - Symptom: When the reranking function is not enabled, it is impossible to understand the relevance of retrieval results to delivery business. Cause: Basic semantic similarity can only match keywords, and cannot distinguish core business metrics such as conversion effects of different materials and budget consumption, leading to relevance judgments that do not meet business needs.
How to Verify Successful Configuration
- Log in to the knowledge base management backend of FastGPT v4.17.0, enter the retrieval configuration page of the corresponding delivery effect optimization knowledge base, and confirm that the
usingReRankswitch is enabled. - Enter a query related to advertising optimization, such as "How to reduce customer acquisition cost", and check whether the retrieval results are sorted by business metrics such as conversions and customer acquisition cost.
- Test the configured reranking model, enter a test text containing advertising data, and confirm that the returned results have no 401 error and conform to the expected sorting rules.
- Check the retrieval log to confirm that the retrieval scope covers advertising data, attribution records, and CRM leads from the past 7 days, matching the preset data source configuration.
The values provided here are common starting points and should be evaluated against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.